Buying a NVDA DGX H100 SuperPOD: $18-20m, plus you've gotta pay for power, cooling, the site, and other upkeep Renting equivalent cloud compute from AWS, Azure: $10-12m per year Renting equivalent compute from a Google TPU v5p Pod: $8.5m per year If you can find cheaper compute, it's probably booked out til 2030. Google will win the AI war.

---

AMD may also play a role in this. Their new repurposed GPU on die accelerators may be sufficient for inference at some scale. Wildly cheaper - but not suited to training, it will bifurcate the AI market into specialized hardware products for for Training and Inference at scale where for now there is only NVDA. The major limitation for widespread training competition that I see after looking into all of this is that NVDA is still the king of memory bandwidth. Those sweet sweet 74% margins are going to go the way of the dodo. It was always going to go this way and there is a lot between the hardware and a usable ecosystem for AI such as good drivers that are very stable. Further, xAi and Tesla's custom silicon is probably going to end up being a blow to demand for NVDA as well. And .. totally anecdotally, I went looking for a high end GPU recently and not only did I have a choice, I also paid MSRP. This hasn't been the case for \*years\* between crypto and AI. I know they said something about the GPU market being robust and blackwell being supply constrained. This may be the case for the DC hardware but it isn't for the consumer and pro cards cards. (Which then makes me wonder what part of the DC hardware is constrained) The DGX also shipped relatively on time and I'd set way met with lack luster demand. Don't get me wrong it looks cool but it's also pretty slow for what it costs.

---

There is an article from the financial times that has done the whole analysis already, with price per vendor, several different chips, etc https://www.ft.com/content/d49707ae-5d6b-473e-9e2b-487d318e6fe9 The DGX A100 cluster cost 25k per chip (since it came with 8 cards and was 200k) Dunno where you got your numbers from

---